Question about Temperature Control

You shouldn't use a delay. However, the system may have a dead time or a delayed response to any control signal.
There are methods for compensating for long dead times. One method is a Smith Predictor. The other is MPC or model predictive control but both methods require a pretty accurate model. To get an accurate model you need to use System Identification to determine the plant mode.
 
Because of the long transport delay or lag times, the controller settings require gradual changes as rapid changes in some system result in over corrections and thermal instabilities.
 
All depends what you are heating up.....a large thermal inertia ....PID control is not good enough and needs embellishments
If I remember correctly it's to do with the dead time vs rising time...i.e. how long it takes to respond to how long it takes to get to 90 percent of its set point.
The real answer lies in the maths, PID loops are relatively simple formulas but the control theory behind it is complex. For large dead time systems, a model is required you need to somehow feedforward the response to the system variations.
A simple trick is to sample the reading very slowly and respond very slowly.....so sample ....respond wait xx time and sample again, by the time you have sampled hopefully something has changed. The down side to this method is you end up with something that is slow and more like a P controller.
 
Never introduce delays into control systems, the faster and better you measure the temperature, with the smallest sensor possible, will give the tightest temperature control.
When you need to control something that is very difficult, and has a long dead-time in the process (i.e. actuator 100% on and the first 15 minutes you see nothing happening to temperature) you need to watch out with the control, since the integrator of the PID is already running to max. and -when- the temperature starts to rises like a rocket and over the setpoint.
Not when you measure fast.
Check our RKC Instrument GZ900: it performs 100 measurements and PID calculations per second. 2 channel version also possible for cascaded (master-slave) control.
See attached the brochure.
 

Attachments

Now if you have thermal mass, warmup times in your heat source, not to mention transport delays of the fluid in the pipe, sensor response time, what ever you do, sort it out first. Fast controller response is meaningless, if the process response time is 20 minutes...
 
Now if you have thermal mass, warmup times in your heat source, not to mention transport delays of the fluid in the pipe, sensor response time, what ever you do, sort it out first. Fast controller response is meaningless, if the process response time is 20 minutes...
exactly
 
Not when you measure fast.
I agree with Dave. Sampling often does not solve the dead time problem.
It is easy to calculate the controller gains that will not cause overshoot if the SOPDT parameters are known.
The formula for the controller gain includes the dead time in the denominator so as the dead time increases the controller gain decreases.
A faster response is possible using a Smith Predictor.

There are formulas for calculating the controller gain and the integrator and derivative time constants even if the dead time is long.
 
The Smith Predictor is not always effective, and a derivative action to speed things up only work in a narrow range of applications where you have servo controls (robotics and related automation), minimal dead time, instant measurements, and minimal time constants, but generally ineffective in most large plant operations where process stability is manditory.
 
The Smith Predictor is not always effective, and a derivative action to speed things up only work in a narrow range of applications where you have servo controls (robotics and related automation), minimal dead time, instant measurements, and minimal time constants, but generally ineffective in most large plant operations where process stability is manditory.
This isn't so. The derivative may 'speed things up' but the real purpose is to the place closed loop poles. If you place the closed loop poles away from the origin then errors will decay faster.
The controller gains are used to closed loop poles. The goal is to place the close loop poles as close a possible on the negative real axis in the s-plane and away from the origin. This will result in faster decay in the error. A SOPDT system has two open loop poles. It takes two gains to place these two open loop poles. The integrator does not count because it comes with its own pole so the proportional and derivative gain are required to move the open loop poles to closed loop pole positions.

Most of the problems people have with the derivative gain comes from bad feed back. The sampling may not be a consistent intervals and the feed back resolution may not be fine enough to calculate accurate rates of change. On top of that there many be noise. All of these limit the usefulness of the derivative gain but...

A Smith Predictor uses a model to calculate the control output. First one must perform a system identification to get the open loop SOPDT parameters. Now the PID is using the model to compute the process variable and rate of change in the process variable. This data is noise free and can be much closer to the actual process variable than the measured process variable.

Too many people are taught this this gain does this and that gain does that etc when in fact they are moving closed loop poles.
The goal should be to move the closed loop poles where you want them.
 
wait until you turn a controller in a heat exchanger loop, it is a multidimensional universe in that case, beyond pole locations, or even more challenging a steam dearator in a boiler feed water loop, with flow, pressure, drum lever, and emergency feed water...every thing is logical and rational except when you are dealing with phase changes, then "look out."
 
wait until you turn a controller in a heat exchanger loop, it is a multidimensional universe in that case, beyond pole locations, or even more challenging a steam dearator in a boiler feed water loop, with flow, pressure, drum lever, and emergency feed water...every thing is logical and rational except when you are dealing with phase changes, then "look out."
A heat exchanger is non-linear, not multi dimensional. The dead time is small enough where a Smith Predictor isn't required. It is easy enough to account for the changing gains of a heat exchanger. Most of the time the heat exchanger is controlling at only one temperature so it is easy enough to assume the gain is linear at that point.

I am familiar with steam generators/boilers. Steam generators do not need Smith predictors either. The sub I was on use mag amps or proportional only control to control the SG water level and feed. I was a reactor controls officer.

There is pretty good info on heat exchanger control at www.controlguru.com. Back in 2005 I contributed to that site.
Note that the control guru has formulas for the FOPDT model. I have derived the formulas for a SOPDT. I use a Taylor series to approximate the dead time.

The real part of the closed loop poles are what cause the errors to decay. You to study pole placement.
 
does pole placement address valve hysterisis?
No, not directly. However, the integrator will compensate if the hysteresis is small. If the hysteresis is a problem then an offset can be added or subtracted from output to compensate depending on the direction of change.

You should look at my YouTube channel, Peter Ponders PID. It is not like most channels. It is very math/control theory intensive.
Most people bail after watching 2 or 3 minutes because they can't keep up. This doesn't bother me as there are lots of YouTube PID channels that cover the basics.
 
Interesting class, root locus plots come in handy too with excel and visual basic, the hysterisis mentioned involved occasional valve stem binding.
 
All those rules and yet the root locus does not provide gains. When it does it provides only one which is OK for governors and similar but not controllers like a PI or PID. There is one more issue that instructors don't mention when they talk about the break away point. That is the point where it is impossible to move all closed loop poles to the left of in the s-plane. This limits the response. Good pole placement should allow ALL closed loop poles to be moved to the left of this breakaway point for a faster response.

The significance of this is that the person doing the control is often asked to do the impossible with an in-adequate controller. There is no way the control person can move all the poles to the left of the break away point unless he has one gain for each closed loop pole.

I am familiar with hysteresis. I have done a lot of hydraulic servo control.

We are getting off topic. Did you see SOPDT1 and the one with the Smith Predictor? Dead time can be compensate for.
 
Didn't watch the vid, but am familiar with the Smith predictor, adaptive controls and various matrix methods for continuous and discrete time systems
 
Top